Chatterbox TTS — CoreML Port
=============================

This repository contains a CoreML conversion of the Chatterbox TTS model
originally released by Resemble AI.

Upstream
--------
  Project:    Chatterbox
  Author:     Resemble AI
  Source:     https://github.com/resemble-ai/chatterbox
  License:    MIT (see LICENSE)
  Copyright:  Copyright (c) 2024 Resemble AI

This port is distributed under the same MIT License. The full upstream
copyright notice and license terms are reproduced verbatim in LICENSE, as
required by the MIT License.

Modifications
-------------
The following changes were made to the original model artifacts to produce
the files in this repository. The modifications cover only the
serialization/inference graph; the trained weights are preserved unchanged.

  * Converted PyTorch / ONNX checkpoints to CoreML `.mlpackage` format using
    Apple `coremltools`.

  * Replaced `complex64` rotary position embeddings (RoPE) inside the s3
    tokenizer with split real `(cos, sin)` tensors. CoreML does not support
    `complex64`; this rewrite is mathematically equivalent.

  * Traced the flow encoder at fixed shapes (400 prompt tokens, 1024 mel
    frames). The host application is expected to chunk longer inputs to
    these limits.

  * Pre-computed and serialized the conditioning tensors for the bundled
    default voice (the five `default_*.bin` files) so callers can run
    inference without re-running the voice encoder pipeline on every cold
    start.

The conversion script (`convert_chatterbox_to_coreml.py`) lives in the host
application's source repository. No model weights were retrained or
fine-tuned; only the serialization format was changed.

Trademarks
----------
"Resemble AI" and "Chatterbox" are referenced here for attribution only and
remain the property of their respective owners. Their inclusion does not
imply endorsement of this port.